We are AlgoCatan, and our project presents a specialized competitive arena designed for benchmarking autonomous Catan agents against both human players and existing Catan bots. By extending the open-source Catanatron framework and game engine, we engineered a high-performance reinforcement learning agent utilizing a combination of self-play, curriculum learning, and intensive hyperparameter tuning. These techniques allowed the bot to iteratively refine its strategic decision-making environment. As a result of our work, our best performing Catan agent successfully outperformed the world’s premier publicly available playable bots, advancing the field of Catan bots.
Team Members:
Jake Read, Sunny Yao, Matthew Cheung, Rebecca Di Filippo
Team Supervisor:
Prof. Istvan David